Kimi K3.1 Pro Weights Drop Nov 4: Licence, Size, How to Run

By Muzamil Ahmad · Updated 2026-10-31 · Coding

#coding#kimi k3.1 pro#open weights#moonshot ai#kimi k3 licence#self-hosted llm#ai model weights

Save

How we write our guides: we start from the vendor’s own announcement or documentation, not from other “best of” lists. Launch dates and features are checked against that primary source; anything reported only by coverage is labelled as reported. Prices and allowances are given only when the vendor publishes them — otherwise we say “not published” rather than guessing. We do not invent test results, prices, or rankings. Our guides are drafted with AI assistance, then reviewed and edited before publication; the final call on what we recommend is a human one.

On November 4, 2026, Moonshot AI is scheduled to do something no lab has done before: hand the public the complete weights of a 2.8-trillion-parameter AI model that independent testing currently ranks as the strongest open model in the world. Kimi K3.1 Pro has been live through Moonshot's apps and API since October 23, but the downloadable release — the one that lets anyone outside Moonshot run, inspect, fine-tune and redistribute the model — is still a promise with a date on it. We checked Moonshot's Hugging Face organisation while writing this article: there is no K3.1 Pro repository yet. The newest thing there is still the older Kimi K3. So this is the guide to read before the drop: what exactly is landing, the licence question that will decide what "open" really means here, the terabyte-scale hardware reality, the two controversies hanging over the release, and a checklist for the day itself. For the launch story — benchmarks, pricing and how the model performs — start with our Kimi K3.1 Pro launch guide; this article is about the weights.

What exactly is dropping on November 4

Kimi K3.1 Pro is not a new architecture. It is a heavily post-trained version of Kimi K3, the model Moonshot released in the summer: a mixture-of-experts system with 2.8 trillion total parameters, of which about 104 billion are active on any given token. It uses Moonshot's Kimi Delta Attention and Attention Residuals design, reads text, images and video in a single model, and works across a context window of one million tokens. What the extra post-training bought is measurable: on Artificial Analysis's Intelligence Index, K3 scored 44; K3.1 Pro scores 55, fourth of every model the evaluator tracks and only three points behind the best closed model, Anthropic's Claude Opus 5.5.

"Full model weights" is the phrase that matters. A weight file is the trained model itself — the numbers that encode everything it learned. With the weights, the model stops being a service you rent from Moonshot and becomes a file you possess: you can run it on hardware you control, fine-tune it on your own data, inspect its behaviour without an API in between, and keep it forever even if Moonshot changes its prices, its terms, or its mind. That is why a weights release is a different event from an API launch, and why the October 23 launch, impressive as it was, is only half the story.

Two things about the drop are still unknown, and both matter. First, Moonshot has not said whether the faster K3.1 Pro Highspeed variant is part of the weights release or stays a hosted-only product. Second — and this is the big one — Moonshot has not named the licence. More on that below, because there is a very good precedent to read in the meantime.

Here is what the family looks like side by side. The K3 column is verified from the actual Hugging Face repository; the K3.1 Pro column is what is known or reasonably expected before November 4, and every expectation is labelled as one.

The licence question — and what K3's licence actually says

Everything about this release hinges on a document that does not exist yet. "Open weights" tells you the file is downloadable; it says nothing about what you are allowed to do with it. That is the licence's job, and Moonshot's announcement named none. The honest position today is: unknown, decided on November 4.

But we are not guessing in the dark, because Kimi K3 shipped under a licence Moonshot wrote for exactly this model family, and you can read it in full in the Kimi K3 repository on Hugging Face. It is a short, unusually readable document, and if K3.1 Pro follows the same pattern — which is an expectation, not a fact — this is what "open" will mean in practice.

The Kimi K3 License starts like the famous MIT licence: permission to use, copy, modify, merge, publish, distribute, sublicense and sell the software — explicitly including the weights, the code and the right to fine-tune and build derivative works — free of charge. For researchers, hobbyists, and the overwhelming majority of companies, that is the whole story: download it and build.

Then come two conditions aimed at the very top of the market. The first is the "Model as a Service" clause. If you or your affiliates run a business that gives third parties API-style access to model inference or fine-tuning — where those third parties meaningfully control inputs, parameters or training data — and that business's aggregate revenue exceeds 20 million US dollars over any consecutive twelve months, you must sign a separate agreement with Moonshot before using the model commercially. Note what the clause deliberately does not cover: products that merely embed the model inside specific features, and simple relays of requests to models hosted by others. The second condition is a branding rule: if the model powers a commercial product with more than 100 million monthly active users, or more than 20 million dollars in monthly revenue, the name "Kimi K3" must be prominently displayed in that product's interface.

A final carve-out softens both conditions: they do not apply to purely internal use — anything that never exposes the model, its outputs or its capabilities to third parties — nor to use through Moonshot's own products and certified inference partners. And the licence closes with the standard "as is" disclaimer and a contact address for licensing questions.

So when the K3.1 Pro licence appears, the checklist is short and specific. Is it the same Kimi K3 License with the name swapped, or a new document? Is the Model-as-a-Service revenue threshold still 20 million dollars? Does the branding rule survive? Has anything been added — an acceptable-use list, a geographic restriction, a research-only tier? Five minutes of reading on November 4 will answer all of it, and until those five minutes happen, any article that tells you what you may do with these weights — including confident claims that they are "fully open" or "not really open" — is getting ahead of the document.

One and a half terabytes: the hardware reality

The second reality check is physical. The Kimi K3 weights on Hugging Face occupy 1.56 terabytes, spread across 96 files, in MXFP4 format — a compact 4-bit representation that Moonshot bakes in during training rather than squeezing down afterwards. K3.1 Pro is the same architecture at the same parameter count, so a similar footprint is the reasonable expectation. To put 1.56 terabytes in perspective: it is roughly twenty-five times the size of a typical 70-billion-parameter model download, and it will not fit on the drive of most workstations, let alone in the memory of a gaming GPU.

Running it is a data-centre conversation, and the arithmetic is not complicated. The weights alone need about 1.6 terabytes of fast memory before you have processed a single token; on top of that sits the key-value cache for that million-token context, which grows with every conversation you serve. A top-end data-centre GPU carries around 141 to 192 gigabytes of high-bandwidth memory. Divide the weights by that and you land in the region of ten to twelve flagship GPUs just to hold the model — one full high-end server node, or a couple of smaller ones linked together — with more headroom needed for comfortable serving. This is not a criticism; it is the honest category of the thing. K3.1 Pro is open in the sense that a power station's blueprints are open: nothing stops you, provided you have a power station.

Three practical consequences follow. First, the first wave of real usage will come from inference providers and well-equipped labs, not individuals — watch for hosted endpoints appearing within days of the drop, the same pattern that followed the Mistral Large 4 open-weights release in late October. Second, community quantisation projects will immediately try to squeeze the model smaller; how much intelligence survives aggressive shrinking of a model that is already 4-bit native is a genuine open question, and early claims should be treated as claims. Third, for almost everyone reading this, the API at $2 per million input tokens and $10 per million output tokens will remain the rational way to use K3.1 Pro for the foreseeable future — the weights matter to you because they cap Moonshot's pricing power and guarantee the model's survival, not because you will host it yourself.

How you will actually run it

If you do have the hardware, the path is already paved, because K3 set it up. Moonshot's K3 release shipped with official support in the two serving stacks that dominate open-model deployment: vLLM, which carries a Moonshot-maintained recipe for the model, and SGLang, which documents it in its cookbook. Day-one support for K3.1 Pro in the same stacks is the reasonable expectation — the architecture is unchanged — but "reasonable expectation" is doing real work in that sentence, and the model card on November 4 is the document that will confirm versions, flags and known issues.

Preparation, if you intend to be ready on the day, is mostly logistics. You need roughly 1.6 terabytes of fast local storage free, a Hugging Face account that has accepted any licence gate the repository carries, and your serving stack installed and tested in advance — K3 itself is the perfect rehearsal download if bandwidth is cheap for you, since its tooling, file layout and deployment guides mirror what is coming. One warning from experience with every major drop: the first hours bring download congestion and a flood of re-uploads, mirrors and repackaged versions from unknown accounts. Download from the official moonshotai organisation only, and let other people be the first to discover which community conversion is broken.

The scoreboard moved: Index v5 and the speed problem

Anyone preparing to evaluate the drop should know that the measuring stick changed last week. On October 27, Artificial Analysis replaced its Intelligence Index with version 5, built mostly on private tests. Under the new index, K3.1 Pro scores 46 — ninth overall, and the highest score of any open-weights model, three points behind the best closed systems on that chart. Some of Artificial Analysis's own task-level measurements are striking: on Terminal-Bench 4.0, the tough terminal-based coding-agent test, they measured K3.1 Pro at 67 percent — the highest they had recorded — and on their GDPval-AA measure of economically valuable knowledge work it ranks third. Moonshot's own reported Terminal-Bench figure, 65.7 percent against Claude Opus 5.5's 66.4, is close behind.

The same independent testing exposes the model's real weakness, and it is not intelligence. It is speed. Artificial Analysis measured K3.1 Pro's output at about 41 tokens per second, ranking 188th of 192 models for output speed. For batch work and offline analysis that hardly matters. For interactive products — a coding assistant answering in real time, an agent loop making hundreds of calls — it matters a great deal, and it is part of why OpenAI's GPT-6.1 Terra, launched October 26 at half the per-token price of the model it replaces and scoring 48 on the new index, is a serious rival for the same workloads. The Highspeed variant exists precisely because Moonshot knows this; whether self-hosting the weights improves or worsens the speed picture will be one of the first things the community benchmarks after November 4. Our comparison of the best AI coding assistants is a useful map of the field this model is entering.

Two controversies hanging over the drop

An honest preview has to include the two stories that broke around this model in the last week, labelled for exactly what they are.

The first is a report, not a finding. On October 28, The Information reported — and Reuters followed on October 29 — that according to people with direct knowledge, Moonshot relied heavily on outputs from OpenAI's GPT-6 Astra to train K3.1 Pro, with a similar claim made about DeepSeek's V4.1 Flash. Training one model on another's outputs, called distillation, sits in contested legal and ethical territory and is restricted by most providers' terms. What is established: the report exists, it is attributed to unnamed sources, and as of October 29 neither Moonshot, OpenAI nor DeepSeek had publicly commented on it. What is not established: whether it is true, and if so, to what degree. It changes nothing about the file that lands on November 4, but it may shape how enterprises' lawyers read the licence — one more reason that document matters.

The second story is Moonshot's own disclosure, and it is stranger: for the open-weight release, the company deliberately made its model worse at hacking. Moonshot says it identified and removed cyberattack-related material from both the mid-training data and the reinforcement-learning tasks, because capability at finding and exploiting software flaws was rising with scale and weights, once published, cannot be recalled. On its internal exploit benchmark, the filtered model scores 81.0 percent; an identical pipeline without the filtering scores 90.5 percent. Ethereum co-founder Vitalik Buterin publicly defended the approach, arguing that restraint now beats the identity-check regimes that a wave of AI-powered attacks could trigger. The open question — and it can only be answered after November 4 — is whether the reduction survives contact with the research community: outside teams will test whether fine-tuning restores the removed capability, and at least one company has already publicly announced plans to try stripping the safeguards. Moonshot's published before-and-after numbers make this the first open release in history whose safety claim arrives with its own test attached.

A third, smaller allegation deserves a sentence for completeness: VentureBeat reported screenshots, which it could not authenticate, purporting to show a paid-promotion brief for Kimi creators that included an instruction not to label posts as paid partnerships. Moonshot had not responded in the reporting. Unverified documents are not evidence — but if you see suspiciously synchronised praise on drop day, you will know the context in which to read it.

Meanwhile, K2 closes its doors today

One dated detail is easy to miss in the drop coverage: the previous generation is being retired on a schedule that starts now. Moonshot's pricing pages indicate the K2 series closes to new users from October 31 — today — and is due to be fully retired by the end of December. If your product still routes traffic to a K2 model, the weights drop is not your only November deadline; migration to K3 or K3.1 Pro belongs on the same calendar page. Moonshot also had a smaller operational stumble this week, apologising for rate limits on K3.1 Pro and resetting a portion of user quotas — a useful reminder that demand for this model, even hosted, is currently outrunning the supply pointed at it.

Your November 4 checklist

When the repository appears, these are the checks that separate knowing from repeating — the same checks we will run before updating this article:

1. The licence file. Name, Model-as-a-Service threshold, branding rule, anything new. Read the document, not the summaries.
2. The repository itself. Official moonshotai organisation, exact repo names — and whether Highspeed is included or hosted-only.
3. Formats and size. MXFP4 as expected? Total terabytes? Any official smaller variants?
4. The model card. Shipped context length, serving-stack versions, known limitations, and any stated difference between the API model and the released weights.
5. Independent numbers. Artificial Analysis and community evaluations run on the released weights — not the API — including whether the cyber-capability reduction is measurable.
6. First endpoints. Which inference providers list it, at what price per token, within the first week.

Conclusion

Strip away the hype and November 4 is a document release as much as a model release. The model itself is already proven in public: fourth on the old independent index, top of the open field on the new one, priced a third below its predecessor, and slow enough that speed — not smarts — is its honest weakness. What nobody outside Moonshot has seen is the paperwork and the files: the licence that decides whether "open" means MIT-free or free-with-conditions, the repository that decides whether Highspeed is included, and the released weights that decide whether the much-discussed cyber-capability reduction is real. Our launch guide covers what K3.1 Pro is; this article will be updated with what the drop actually contained — licence first — as soon as the repository is live and verified. If Moonshot keeps its date, that update lands on November 4.